Re: [PATCH 1/3] cgroup/rstat: Fix forceidle time in cpu.stat

[Abel Wu <wuyun.abel@bytedance.com>]

On 1/24/25 5:22 PM, Michal Koutný Wrote:
> Hello.
> 
> On Fri, Jan 24, 2025 at 01:47:01AM +0800, Abel Wu <wuyun.abel@bytedance.com> wrote:
>> The commit b824766504e4 ("cgroup/rstat: add force idle show helper")
>> retrieves forceidle_time outside cgroup_rstat_lock for non-root cgroups
>> which can be potentially inconsistent with other stats.
>>
>> Rather than reverting that commit, fix it in a way that retains the
>> effort of cleaning up the ifdef-messes.
> 
> Sorry, I'm blind, where's the change moving wrt cgroup_rstat_lock?
> (I only see unuse of root cgroup's bstat and a few renames).

Hi Michal,

The following hunk deleted the snapshot of cgrp->bstat.forceidle_sum:

  	if (cgroup_parent(cgrp)) {
  		cgroup_rstat_flush_hold(cgrp);
  		usage = cgrp->bstat.cputime.sum_exec_runtime;
  		cputime_adjust(&cgrp->bstat.cputime, &cgrp->prev_cputime,
  			       &utime, &stime);
-#ifdef CONFIG_SCHED_CORE
-		forceidle_time = cgrp->bstat.forceidle_sum;
-#endif
  		cgroup_rstat_flush_release(cgrp);
  	} else {

and then read forceidle_sum from @cgrp directly outside of the lock,
but its value can be changed in this window, so...
  
-#ifdef CONFIG_SCHED_CORE
-	seq_printf(seq, "core_sched.force_idle_usec %llu\n", forceidle_time);
-#endif
+	cgroup_force_idle_show(seq, &cgrp->bstat);
  }
  
result in the inconsistence between forceidle and other cputimes.

Best Regards,
	Abel